Uncertainty-aware deep learning with physics-informed bayesian sampling for lithium-ion battery prognostics

Abstract

Accurate state-of-health (SOH) estimation and remaining useful life (RUL) prediction are critical for the reliability, longevity, and safety of lithium-ion battery systems. While data-driven methods have advanced battery prognostics, most struggle with dynamic operating conditions, heterogeneous degradation patterns, and limited ability to provide reliable uncertainty quantification (UQ). To address these challenges, we propose DG-PNUTS, an uncertainty-aware deep learning framework that integrates dual gated recurrent unit (GRU) networks with a physics-informed Bayesian No-U-Turn Sampler (NUTS). NUTS is an adaptive Markov Chain Monte Carlo (MCMC) algorithm that enables principled posterior inference and UQ. The framework employs a divide-and-conquer strategy by training multiple GRUs on subgroups of aged battery data, effectively capturing heterogeneity. For in-service batteries with limited historical data, Bayesian multi-source domain adaptation transfers knowledge from pre-trained models, with the physics-informed NUTS enhancing inference and reliability. A standalone GRU is further utilized for RUL prediction based on the estimated SOH and extracted health indicators. The proposed method was validated on multiple publicly available accelerated aging datasets, demonstrating superior accuracy, robustness across varying operating conditions and chemistries, and reliable UQ compared to benchmark methods. These results highlight the effectiveness of combining deep learning with physics-informed Bayesian MCMC sampling for uncertainty-aware battery prognostics.